Standalone libgomp with MIC backend for explicit CUDA-like Xeon Phi device programming

This is a port of libgomp and liboffloadmic libraries for standalone use on Intel Xeon Phi K1OM devices. The codebase has been extracted from GCC 5.3.0 source tree. Central Makefile supports building for native Xeon Phi device or host emulation. As a prerequisite, building environment must have access to GCC crosscompiler for MIC (e.g. /usr/linux-k1om-4.7/bin/x86_64-k1om-linux-gcc shipped with Intel MPSS).

Features

CUDA-like API provided in libmicrt library enables direct user access to MIC devices enumeration, memory transfers and device kernels launching:

micError_t micGetPlatformName(const char** name);

micError_t micGetDeviceCount(int* count);

micError_t micSetDevice(int device);

micError_t micMalloc(void** devPtr, size_t size);

micError_t micFree(void* devPtr);

micError_t micMemcpy(void* dst, const void* src, size_t size, enum micMemcpyKind kind);

micError_t micLaunchKernel(const char *funcname, void *args);

Building for Native Xeon Phi

$ git clone https://github.com/apc-llc/liboffloadmic.git
$ cd liboffloadmic
$ git submodule init
$ git submodule update
$ make target=native

Building for Host Emulation

$ git clone https://github.com/apc-llc/liboffloadmic.git
$ cd liboffloadmic
$ git submodule init
$ git submodule update
$ make target=emul

Testing

Host emulation:

$ cd liboffloadmic/test
$ make
$ make run
LD_LIBRARY_PATH=:/home/marcusmae/apc/liboffloadmic/test/../install/target/lib/../lib ./test
CUDA-like runtime API powered by GOMP backend for intelmic
1 intelmic device(s) available
num_threads = 8
Result is correct

Xeon Phi 5110P:

$ cd liboffloadmic/test
$ make
$ make run
SINK_LD_LIBRARY_PATH=:/home_lustre/dmikushin/liboffloadmic/tests/saxpy/../../install/target/lib/../lib64:/home_lustre/dmikushin/liboffloadmic/tests/saxpy/../../install/../tbb/build/linux_release ./test
CUDA-like runtime API powered by GOMP backend for intelmic
1 intelmic device(s) available
num_threads = 236
Result is correct